Upgrade to remove ads
Only $35.99/year

Lecture 7 - Visual Object Detection

Terms in this set (17)

What is the difference between Object recognition task and Object detection task?
Object recognition task:
- Is there a dog or a sofa in the image?

Object detection task:
- Where are the dog and the sofa in the image?
List some ways of evaluate object detection
1) >= 0.5 Intersection over Union (IoU)

2) Precision
#True detections/#Detections

3) Recall
#True detections/#Positive examples
What are the classical steps in a Object detection process?
1) Object Model Specification
2) Hypothesis Generation
3) Hypothesis description and scoring
4) Detection Refinement
List some Object models
1) Statistical template in bounding box
- Template visualization is used for bounding box
- Object (x, y, w, h) in image

2) Articulated parts model
- Object is configuration of parts (arm-box, leg-box etc)
- Each part can be detected

3) Hybrid model:
- Use both template and parts
List the different methods for hypothesis generation
1) Sliding window
2) Region proposal
How does sliding window works for hypothesis generation?
1) A window is sliding over the image
2) Each window is separately classified
How does region proposal work for hypothesis generation?
Find "blobby" image regions that are likely to contain objects
What are some positive and negative aspects with sliding window (hypothesis generation)?
+ Simple
+ Repeatable
- Exhaustive and slow
What are some positive and negative aspects with region proposal (hypothesis generation)?
+ Fast execution
+ Fewer candidate boxes/regions
- Random, not repeatable
- Requires preprocessing
List some examples of hypothesis description and scoring methods
1) HOG based template
2) HOG based template + parts
3) CNN features
How does the Dalal-Triggs Pedestrian Detector work?
1) Input image
2) Normalize gamma & colour
3) Compute gradients
4) Divide into cells and compute magnitude and angle
5) Normalize blocks of cells
6) Collect HOG's over detection window (concatenate each block into a feature vector)
7) Feed the feature vector into a SVM
8) Classify if person or not
How does Deformable part based detectors work?
• It defines objects by a collection of parts models by:
1) Appearance
2) Spatial configuration of parts (how they are spatial position relative to each other)
How does R-CNN works?
• It uses Region proposal + CNNs

1) Inputs an image
2) Extract region proposals of the image (~2k)
3) Feed region proposals into CNN
4) CNN classifies region
How many features (cascade layers) does a real-time Viola-Jones detector use?
36 layers
How does the classifier in Viola-Jones Face detection work?
It's a linear combination of weak classifiers that are optimized using Adaboost
Recommended textbook explanations

Computer Organization and Design MIPS Edition: The Hardware/Software Interface

5th EditionDavid A. Patterson, John L. Hennessy
220 explanations

Introduction to Algorithms

3rd EditionCharles E. Leiserson, Clifford Stein, Ronald L. Rivest, Thomas H. Cormen
709 explanations

Introduction to the Theory of Computation

3rd EditionMichael Sipser
389 explanations

Operating System Concepts

9th EditionAbraham Silberschatz, Greg Gagne, Peter B. Galvin
483 explanations